Move from dashboards to decisions
I design analytics that connect model outputs to field actions: what changed, why it matters, and what a service or quality team should do next.
Featured Project
Predictive Maintenance on NASA CMAPSS
Open-source implementation
predictive-maintenance-cmapss
An end-to-end Python pipeline on NASA's CMAPSS turbofan degradation dataset — strict-schema data loader, feature engineering, baseline and gradient-boosted RUL regressors, with executed Jupyter notebooks showing every result.
96%test coverage
59tests passing
3.11–3.13Python matrix CI
MITlicensed
Python
scikit-learn
XGBoost
pandas
Plotly
Jupyter
GitHub Actions
uv
ruff / mypy
Results from the executed notebooks
Real benchmarks on FD001
Every figure below is rendered straight from the executed notebook in the repository — click any card to open the full notebook on GitHub.
EDA — trajectories aligned to failure
25 training units, sensor 11 plotted vs. cycles before failure. Clean monotonic drift in the last ~80 cycles motivates the piecewise-linear RUL relabelling.
RUL distribution — raw vs. clipped
Capping the regression target at 125 cycles concentrates model capacity on the regime where degradation is observable. Heimes (2008) convention.
Ridge baseline — RMSE 18.27 / S-score 592.6
Standard-scaled L2 regression on rolling features. Sets the floor that any non-linear model must clearly beat to justify its complexity.
XGBoost head-to-head — RMSE 18.23 / S-score 814.8
Tuned XGBoost edges out Ridge on RMSE but loses on the asymmetric S-score. The honest result: FD001's single regime is exactly where linear features compete with trees.
Feature importance corroborates the EDA
Rolling statistics on high-pressure-compressor sensors dominate — exactly the channels whose drift was visible in the EDA notebook.
Coming next
FD002 / FD004 multi-regime evaluation
Operating-regime clustering, regime-aware normalisation, and an LSTM sequence model over full trajectories — the setting where XGBoost is expected to clearly win.
Track progress on GitHub →Interactive concept preview
Synthetic dashboard — live in your browser
A self-contained Plotly visualisation of what the same pipeline looks like running against real-time sensor streams. The data is synthetic so the page stays static — for the actual benchmark numbers, see the cards above.
Multi-Sensor Time Series — Degradation Monitoring
RUL Prediction — Actual vs. Predicted
Anomaly Detection — Operating State Classification
Equipment Health Score (Current)
Concept summary
| Sensors | Vibration, Temperature, Pressure |
| Cycles | 300 operating cycles |
| Alert threshold | Health score < 30% |
| Implementation | eastani/predictive-maintenance-cmapss ↗ |